Το work with title Comparison of artificial intelligence systems for the detection of objects on UAV-based images by Trimas Christos is licensed under Creative Commons Attribution 4.0 International
Bibliographic Citation
Christos Trimas, "Comparison of artificial intelligence systems for the detection of objects on UAV-based images", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2022
https://doi.org/10.26233/heallink.tuc.93339
Unmanned Aerial Vehicles (UAVs) have experienced great growth and as of 2020 at least 100 countries use UAVs in tactical options, while at the same time even more commercial applications deploy drones, for example photography and filmmaking, smart crops, smart cities, emergency handling, drug delivery, traffic management, etc. The big success of UAVs came due to the huge growth of electronics and the revolution of data. One of the most popular application of drones is object detection before designing the planned operation, e.g. differentiate pedestrians from cars or bikes in cross-road management systems. Deep Learning algorithms have been proven to be the best solution in such kind of problems. This diploma thesis collects and studies some of the most well-known detection systems, it analyzes in theory and in practice an object detector, the famous single-stage detector RetinaNet. Furthermore, a modified model is proposed that utilizes more Convolutional Blocks and combines features from different levels of the Neural Network. The extra convolution block is a mirror of the FPN network; therefore, the new model is called “Two-Phase Feature Pyramid Network Retina”. Since the goal is to compare those models, the classic RetinaNet and the modified model, were trained and tested using the Stanford Drone Dataset, a dataset designed to train object detectors for UAVs. The modified model achieves an accuracy score 6% higher than the baseline model, and it seems to outperform the original model in every metric, such as Precision, Sensitivity and F1 score. Finally, both the original and the modified Retina, were compared with other well-known object detectors such as YOLO, Faster RCNN, SSD, etc. The proposed architecture seems to outperform almost every object detector from the literature in terms of mean Average Precision. In conclusion, the modified model can be used to detect small objects in applications where accuracy is a critical factor.